How to use with pytorch 1.0 for compatible with CUDA 10?

[sangnguyenz]

Is is possible to train faster-rcnn with pytorch 1.0 in this project?
Thank you.

[timmeinhardt]

Unfortunately the underlying faster-rcnn repository does not work with PyTorch 1.0. However, the most recent version of torchvision provides Faster R-CNN out of the box and it should be quite easy to use our Tracktor code with that version of a Faster R-CNN. This would definitely be a contribution that we are happy to integrate into our repository.

[sangnguyenz]

Unfortunately the underlying faster-rcnn repository does not work with PyTorch 1.0. However, the most recent version of torchvision provides Faster R-CNN out of the box and it should be quite easy to use our Tracktor code with that version of a Faster R-CNN. This would definitely be a contribution that we are happy to integrate into our repository.

Thank you for your reply. I've done taking boxes and classes from FRCNN pre-trained model from PyTorch as you said. Next, Where can I take the code for integrating the detector and tracktor? Do I need to modify the tracktor evaluation for demo code?

Thank you

[timmeinhardt]

You would have to modify multiple files. Definitely experiments/scripts/test_tracktor.py and src/tracktor/tracker.py. It is definitely possible to switch out the underlying Faster-RCNN implementation but there is probably some work involved.

[sangnguyenz]

You would have to modify multiple files. Definitely experiments/scripts/test_tracktor.py and src/tracktor/tracker.py. It is definitely possible to switch out the underlying Faster-RCNN implementation but there is probably some work involved.

Excuse me, Can you help me to explain the differences between Tracker and Oracle Tracker?
Thank you in advance.

[timmeinhardt]

Please have a look at Section 4.2 of our paper. There, we explain the purpose of our oracle trackers and their distinctions.

[sangnguyenz]

You would have to modify multiple files. Definitely experiments/scripts/test_tracktor.py and src/tracktor/tracker.py. It is definitely possible to switch out the underlying Faster-RCNN implementation but there is probably some work involved.

After taking a look at those 2 files, I'm not sure that I understand how you can get the data to put in the network. Now I'm working with Pytorch 1.0, so I can not produce blob data as the way you did. Can I put the frame as image or image array to tracker.step(frame) of the tracker (supposed that I've done get detection step).?

Thank you.

[timmeinhardt]

Without testing PyTorch 1.0 by myself I think that the data blobs should still be produced as before, i.e, the Tracker.step method receives a blob argument which includes the data, ground truth and some image information.
In order to use a different object detector (our detector will not work with PyTorch 1.0), you have to provide the same interface as ours. See self.obj_detect.load_image and self.obj_detect.test_rois and the corresponding functions in src/tracktor/fpn.py.

[idmakers]

Hello ! I already have a detect data , and I want to use my own data .How can I do it?

[timmeinhardt]

@idmakers Please open an issue for each problem. This one is only about the PyTroch 1.X support of this repository. I answered your question in a few previous already closed issues.

[timmeinhardt]

We just updated the repository to support PyTorch 1.3 and added a new object detector. I did some rudimentary testing but feel free to give me feedback if something is not working as before.

I will close this issue. Please open new issues if something is broken now.